Robot bin picking: 3D pose retrieval based on Point Cloud Library

This work covers the problem of object recognition and 6 DOF pose estimation in a point cloud data structure, using PCL (Point Cloud Library). The result of the computation will be used for bin picking purposes, but it can also be applied to any context that require to find and align a specific pattern. The goal is to align an object model to all the visible instances of it in an input cloud. The algorithm that will be presented is based on local geometry FPFH descriptors that are computed on a set of uniform keypoints of the point clouds. Correspondences (best match) between such features will be filtered with RANSAC procedure: from this data comes a rough alignment, that will be refined by ICP algorithm. Robust dedicated validation functions will guide the entire process with a greedy approach. Parallelism has also been implemented using OpenMP API. Time and effectiveness will be deeply discussed, since the target industrial application imposes strict constraints of performance and robustness. The result of the proposed solution is really appreciable, since the algorithm is able to recognize almost all the present objects, with a minimal percentage of false negatives and an almost zero false positives rate. Experiments have been conducted on a large dataset, that was acquired with a triangulation system made up by one camera and two intersecting lasers as structured light sources. Such vision system has been mounted first on a fixed position over a conveyor belt, then on a moving robotic arm, in order to cover a larger are